Criterica Intelligence — production models trained on real court records, not synthetic data
Case Study · Deidentified

What the audit found in a $70M book.

A national pre-settlement litigation finance fund engaged Criterica to run a Portfolio Intelligence Audit against its active book. We scored roughly 8,400 cases with the Criterica production model fleet, trained on real court records. What follows is the finding, with the firm and every counterparty anonymized and figures stated in bands.

"The fund did not have a returns problem. It had a visibility problem. Every finding below was already inside the portfolio. The audit made it legible, and put a number on what better intake scoring, pricing, and monitoring would have changed."

The engagement
~$70M
Active book audited
~8,400
Active cases scanned
~3x
Blended MOIC (reported)
~44%
Of book carried no risk score
The findings

Five things the fund did not know about its own book.

01
Operational system gaps

A reported return that the data did not support

The fund's case-management system reset each firm's lifetime funded amount to zero when a case closed. It tracked outstanding principal, not capital deployed. The largest relationship therefore appeared to return above 4x. Rebuilt against lifetime capital, the real figure was closer to 1.6x. The reporting was not wrong on purpose. The data architecture made the overstatement invisible.

02
Concentration & performance bands

The best returns sat in the smallest positions

The largest firm relationship, roughly $100M deployed over the life of the program, returned in the mid-1x range. The highest-returning relationship was a sub-$300K position compounding above 10x. Capital was concentrated where returns were lowest. The audit named the inversion the fund could not see from its own dashboards.

03
Monitoring failures

Nearly half the book could not be screened

About 44% of the active book, on the order of $31M, carried no risk classification and no expected-recovery amount. Those positions could not be scored, monitored, or stress-tested. The single largest source of unmanaged variance was not a bad case. It was a missing field at intake.

04
Capital velocity

Idle capital was quietly costing return

Settlements were redeploying with roughly a 45-day lag, leaving on the order of $9M sitting idle at any moment. That drag translated to approximately $325K per quarter in foregone return. Tightening the redeployment cycle toward 15 days recovered six figures annually, without writing a single new case.

05
Early warning

Deteriorating relationships were visible early

Firm-level outcome signals flagged do-not-use relationships 6 to 18 months before a manual review would have caught them. Two relationships already showing negative expected return were still receiving capital. Early warning is the difference between a managed exit and a write-down.

Quantified impact

What acting on the findings would have been worth.

$325K/quarter

In foregone return from a 45-day redeployment lag. Tightening to 15 days recovers it without writing a new case.

44% of the book

Carried no risk score. Positions that cannot be monitored cannot be managed. Single largest source of uncontrolled variance.

6–18 months early

Advance warning on deteriorating firm relationships — the difference between a managed exit and a write-down.

On variance

The audit does not claim every variance it surfaced was an avoidable loss. Some of those matters were still profitable. The value was identifying which variances were foreseeable through better intake scoring, pricing, duration modeling, concentration controls, and monitoring, and quantifying what catching them earlier would have been worth.

From audit to monitoring

The one-time finding became a live system.

After the audit, the same models run continuously against the active book. What was a snapshot becomes an alerting layer.

01Live MOIC and IRR tracking by firm, case type, and vintage
02Settlement and redeployment forecasting across the active book
03Do-not-use early warning on deteriorating firm relationships
04Underwater-advance screening against jurisdiction medians
05Capital allocation intelligence and concentration thresholds
06Docket and outcome monitoring on every active matter
See what the audit finds in your book.
Bring your portfolio. We run the models against your own data.
Audit Methodology
Data received

A funder-provided export of the historical and active book — matter attributes, funding amounts, servicing records, and resolved outcomes where available.

Reconciliation

Matters were reconciled against servicing records before analysis; unmatched or incomplete records were flagged and excluded from affected calculations rather than imputed.

Model application

Production models were applied to matter attributes as recorded at intake. Scores reflect what the models would have said at the decision point, not with hindsight.

Observed vs. modeled

Capital and classification figures (including the unscored share and idle-capital finding) were observed directly from the funder's own records. Forward-looking figures are modeled and labeled as such.

Nature of findings

Findings are retrospective associations identified in the book as it stood — they are diagnostic, not causal claims, and not projections of future performance.

Limitations

Results reflect one portfolio, its intake filters, and its record quality. Selection effects apply. A different book will produce different findings — which is the point of running the audit on yours.

This case study is published with the client's identity and all counterparties removed. Financial figures are stated in approximate bands rather than exact values. It is presented to illustrate the Portfolio Intelligence Audit methodology, not as a forward-looking projection of returns.